@vineetdaniel
Account based inIndia
About this account
- Account based in
- India
- Connected via
- India App Store
Account-level information from X, not a live location or the device used for a specific post.
Building AI & Tech products at https://nitter.cf/t.co/hX3iwpvWKV
India
Joined March 2009
- Tweets1.9K
- Following1.2K
- Followers311
- Likes673
Vineet Daniel विनीत डैनियल retweeted
Run Bonsai 27B locally on a 16 GB Mac 🥷
Bonsai 2 27B is @PrismML's ternary build of Qwen3.8 27B that keeps 98.2% of FP16 quality in 7 GB, it made a voxel Japanese pagoda in one prompt with Three.js!
Run AI models locally - atomic.chat
Today, we’re announcing Ternary Bonsai 2 27B.
Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance.
Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use.
Ternary Bonsai 2 27B is available today under Apache 2.0.
How to use Jev, and where it actually gives you the 100x:
setup takes 10 minutes:
1. join the waitlist, people are getting approved same day
🔗 typesafe.ai
2. install the official skill so your agent writes correct calls:
- npx skills add typesafe-ai/skills --skill typesafe-ai
on Claude Code it's two commands, the marketplace add on its own doesn't install anything:
- claude plugin marketplace add typesafe-ai/skills
- claude plugin install typesafe@typesafe-ai
3. create an API key in the dashboard
4. in your prompt just say: "use the TypeSafe skill"
now the part nobody is posting:
the 100x isn't the model, it's where you put it
you don't get it by swapping your LLM for Jev
you get it by deleting the calls that never needed a language model
open your agent and find every call that just picks something:
> which tool next
> is this spam
> is this chunk relevant
> does this need a human
> is this diff risky
none of those are writing tasks
they're if statements you outsourced to a frontier model
here's the upgrade, in order:
1. replace each one with a typed question
Choice picks from up to 255 options, Score places it on a 2-10 level scale, Noul returns a raw 0-1
2. batch them
questions in one call run in parallel and barely move the latency, and output tokens are free
so ask every question you might need, including the ones you'll throw away
3. threshold on confidence, not on the answer
under 0.5 escalate to a big model or a human
0.85+ before anything irreversible
4. never let it invent options
build the candidate list in code, from the DOM, the retriever, the tool trace
then let it pick
5. put it in the loop, not next to it
router picks the cheap model, gate checks the tool call before it runs, judge verifies the output after
that's where the heaviest calls in your agent are hiding
6. start with compaction tonight
score every tool call, drop the dead ones, keep the survivors verbatim instead of a lossy summary
lowest effort win available and you'll see it on tomorrow's bill
the honest part:
text only right now, no images, no audio
and on broad benchmarks it loses to frontier models
but somebody ran 18,514 emails through it zero-shot and got 98.33%
against a TF-IDF classifier trained on 14,800 labelled examples that got 98.39%
no training data, $1.12 total
it wins on narrow, well specified decisions
which is most of what your agent is actually doing all day
today gonna share use case how i integrated it to content creation and how i find winning meta ads now in a seconds...
found the perfect use case for @typesafeai Jev:
instant compaction
in 2026, why is compaction still a summarization prompt?
Jev can make it instant by scoring every tool call and dropping what’s irrelevant
Requsting builders to also share how much they paid in AI bills alongwith statements like AI running entire company, swarm of agents, developed SaaS product in two days etc. Cost gives real picture. I started with grok bots.
and ended up exhausting weekly limits in a day. Then had to upgrade cursor subscription to know that it doesn't even change weekly subs. Added 10$ to continue my work and it got wiped in an hour. Waiting for weekly limits to reset.
@vineetdaniel Introducing iPhone Duo, iPhone 18 Pro, Apple Watch Series 12, Apple Watch Ultra 4 and AirPods 5.
Vineet Daniel विनीत डैनियल retweeted
We benchmarked ourselves against Ollama and llama.cpp on one idle M2 Pro mini, then published the raw JSON.
Decoded tok/s, 8 streams, Qwen3.6-35B-A3B:
rapid-mlx 82.9
llama.cpp 48.8
Ollama 27.2
llama.cpp prefills cold prompt-batches faster on a dense 12B. Count prefill in the wall clock and that 3× becomes 1.6×. We published those tables too.
Run it on your own Mac.
Vineet Daniel विनीत डैनियल retweeted
Introducing GlucoFM, a lightweight, self-supervised continuous glucose monitoring foundation model that separates metabolic baselines from transient spikes, producing transferable representations and setting new performance standards across diverse metabolic prediction tasks, such as diabetes risk assessment, insulin resistance, and post-prandial glycemic response.
Learn more →goo.gle/4cbfigr
ALT Overview of GlucoFM, a lightweight dual-stream foundation model for continuous glucose monitoring.
@pvvnlmeerut no power supply in netaji nagar since 2am. No one picking up calls and local office is locked. Please look into the issue on priority.
@1912PVVNL please look into it..no power supply in netaji nagar, meerut area since 2 AM.
Germany has launched one of the world's best open-source AI models.
Soofi S, made by the Soofi consortium, is a 30B parameter model fully trained in Europe and tops the ranking for open-source AI.
Huge moment for Europe, and finally some competition for Chinese open-source AI.
Vineet Daniel विनीत डैनियल retweeted
Big news: The GLM-5.2 Colibri model is here, and it's a game changer for multilingual AI. This int4 quantized MoE model runs efficiently on CPUs, making advanced language tech more accessible.
Vineet Daniel विनीत डैनियल retweeted
وداعًا لـ vLLM و الـ GPUs المليونية.. وحش الـ Open Source الجديد يشتغل على جهازك بـ 25 جيجا رام بس! 🤯💻🔥
مشروع colibri يقلب موازين تشغيل الموديلات العملاقة محليًا، حيث تمكن من تشغيل الموديل الضخم GLM-5.2 (ببنية MoE وحجم 744 مليار باراميتر) على ذاكرة عشوائية 25 جيجابايت فقط وبدون الحاجة لبطاقات رسومية (GPUs) باهظة الثمن!
🎯 الكود مكتوب بالكامل بلغة C صافية وبدون أي اعتمادات ثقيلة، مما يعطي أقصى كفاءة وسرعة استجابة للمكونات.
💾 بدلاً من تحميل الـ 744B باراميتر بالكامل في الرام، الأداة تقوم باستدعاء وبث الموديلات الفرعية (Experts) المطلوبة فقط من الهارد ديسك (SSD) إلى الرام لحظة بلحظة أثناء المعالجة، مما يحجم استهلاك الذاكرة بشكل مرعب!
احفظوا التغريدة (Bookmark) 📌 لتجربة المستودع على أجهزتكم، ورابط الـ Repo المباشر بالأسفل! 🚀💻👇
Vineet Daniel विनीत डैनियल retweeted
744B parameters. On a laptop. With 25GB RAM.
Colibri runs GLM-5.2 (744B MoE) in pure C with zero dependencies. The trick: only ~40B params activate per token, so it keeps the dense part resident and streams experts from disk on demand. A single 2,400-line C file. No GPU, no BLAS, no Python at runtime.
This shouldn't work. But it does.
⭐ 2.1K #AI #OpenSource
github.com/JustVugg/colibri
Follow for daily dev finds 🔔
Vineet Daniel विनीत डैनियल retweeted
Local AI was never this EASY
> Install ODS
> Let it detect your hardware
> It will download the best model for your hardware
> And then start local inference and Open WebUI for you
With ODS, you can
> Add voice, agents like Hermes, workflows, RAG, search, image generation, and more
> Manage the whole stack from one dashboard
Now your PC, Mac, or Linux box is a private AI server
No cloud required
No subscription required
Your prompts and data stay on your machine unless you choose otherwise
We're gonna make Local AI The Default
Local AI 101 - Models by Parameter Size
The 'B' stands for Billion parameters. A good rule of thumb is 1B = roughly 1GB of VRAM.
~1B (Nano): Used for specific simple tasks like OCR, or built into web apps for lightweight execution.
~9B (Micro): Convenient as a support model running on small hardware like smartphones or laptops.
~35B (Mini): The sweet spot for standard consumer devices. Once fine-tuned, they perform specific tasks very well.
~400B (Haiku size): Includes lightweight models like Deepseek V4 Flash and Minimax M3. The max size runnable on a single 128GB Mac using dynamic quant.
~800B (Sonnet size): Mid-size models like GLM-5.2 and Kimi. Requires 512GB+ VRAM, but excellent for complex tasks like coding.
1.6T+ (Opus size): Massive models like Deepseek V4 Pro and Longcat. Too heavy for consumer devices; requires server-grade hardware.
open.substack.com/pub/vineet…
Another weekend, another poc.
calc-liard-seven.vercel.app/ check what you need to run AI locally. No more guess work. Try and share the feedback.
#localai #artificialintelligence #localforge
Vineet Daniel विनीत डैनियल retweeted
Good morning y'all!
Qwopus-3.6-35B-A3B-MTP-Coder is live! All GGUF's will be populating over the next few hours!
It's a lightning-fast MOE with the coder curriculum recipe. Similar to the 27B coder, it shines with thinking disabled, offering significantly faster wall time for similar, and in some cases superior results to same-sized thinking alternatives! With thinking disabled, it goes toe-to-toe with the new Ornith 35B MoE across a huge eval suite (performed by @no_stp_on_snek), edging it on the coding trajectories and decisively on speed and cost, even though Ornith was run with thinking enabled.
See the model card for the full test results, and shoutout to Tom, @no_stp_on_snek, for thoroughly evaluating the model for us before launch!
With MTP and thinking disabled, along with the MOE speed, it runs so quickly in harnesses like @opencode that it almost feels instant @ 253 tps on my 5090.
No 8k tokens of thinking before a coherent output is actioned. This is especially useful in long contexts, where the base models will progressively start thinking for tens of thousands of tokens before replying.
Compared to the base models with thinking off, the coder curriculum really advances the no-think frontier. Especially in terms of how creative it can be. Run temp hot as usual, 0.85-1, and make sure your harness isn't overriding the temp setting of your server at runtime.
If you want to use it to its full ability, I would recommend giving it very thorough prompts. I have been using it in opencode, and I have been blown away by the results it generates autonomously with chunky prompts. Please see links to the demo's Aether Dominion (RTS Game), and a slide deck presentation the model made about itself that turned out beautifully, links in comments below!
I am getting results on this incredibly fast local model (with thinking disabled) that I couldn't get in some thinking frontier models over a year ago.
Open source is accelerating fast, and in light of recent events, there's never been a better time to get your local AI workflows tightened up. This MOE would be a great one to play with, and it's also a great one if you don't have much VRAM because it can run fast offloaded partially to system memory!
All of that said, please give it a run with thinking off and build something you'd like to see. We'd love to see your results and any feedback on specific use cases in the comments below!
Also, thanks so much for 5k followers, you all make up such an enjoyable and knowledgeable open source community, and I am so blessed to be able to collaborate and discuss this research with all of you. I can't express how grateful I am for every comment. As always, I will try to reply to them all!
If we ever get monetized on X, I will put every penny into buying more hardware for our lab!
Have a blessed day, my friends, looking forward to your thoughts!
huggingface.co/Jackrong/Qwop…
Vineet Daniel विनीत डैनियल retweeted
Big news for AI builders: NVIDIA dropped a quantized version of Qwen3.6 that fits in 35B parameters but runs like a 3B model. This FP4 MoE beast is a game changer for efficient inference.
Vineet Daniel विनीत डैनियल retweeted
📣📣 Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc adaptation.
🤔 LLMs are trained to be better agents — better at acting in environments. But nobody has trained them to model the environments themselves.
🗺️ Our roadmap: investigate how language world modeling can push the boundaries of general agent capabilities, along two routes:
1️⃣ Build a foundation model for environment simulation — outperforming Claude Opus 4.8 and GPT-5.4 on AgentWorldBench
2️⃣ Investigate how world modeling enhances agent training:
🔬 Controllable Sim RL (agentic RL with LWM as environments) surpasses training in real environments
🧠 Learning to predict environments (LWM warm-up) makes agents stronger — remarkably, even without any agent-specific training, this predictive knowledge transfers to agentic tasks with zero fine-tuning
📑 Paper: arxiv.org/abs/2606.24597
📖 Blog: qwen.ai/blog?id=qwen-agentwo…
💻 GitHub: github.com/QwenLM/Qwen-Agent…
🤗 HuggingFace: huggingface.co/collections/Q…
🧩 ModelScope: modelscope.cn/collections/Qw…
